Causal Mediation Analysis in Politics
Also known as: Causal mediation, Mechanism analysis, Direct and indirect effects, Potential-outcomes mediation
Causal mediation analysis decomposes the effect of a treatment — often a randomized experimental manipulation, such as a campaign message or an information treatment — into the part transmitted through a specified intermediate variable, the mediator, and the part operating through all other pathways. Formalized in the potential-outcomes framework by Imai, Keele, Tingley, and Yamamoto, it defines the average causal mediation effect (ACME) and the average direct effect, makes explicit the sequential-ignorability assumption required to identify them, and supplies a sensitivity analysis for when that assumption fails. It lets political scientists move beyond 'does the treatment work?' to 'why does it work?'
Key highlights
- Defines direct and indirect effects rigorously in the potential-outcomes framework, separating mechanism from total effect.
- Applies to nonlinear models and interactions, not just linear structural-equation coefficient products.
- Provides a principled sensitivity analysis that quantifies how much hidden confounding would overturn the result.
- Connects experimental treatment effects to substantive theories of why a treatment works, advancing mechanism testing.
Intuition
This section is available to Pro members. Upgrade to Pro
How it works
This section is available to Pro members. Upgrade to Pro
When to use it
Use causal mediation analysis when you have a treatment, a measured intermediate variable hypothesized to carry its effect, and an outcome, and you want to quantify how much of the effect flows through that mechanism. It is most credible when treatment is randomized, as in survey or field experiments, and when you can defend the assumption that the mediator is not confounded with the outcome conditional on covariates. It is less appropriate when multiple mediators operate jointly, when the mediator is strongly confounded with the outcome and no sensitivity analysis can rescue the estimate, or when treatment-mediator interactions are complex and the design provides no leverage on them.
Strengths & limitations
- Defines direct and indirect effects rigorously in the potential-outcomes framework, separating mechanism from total effect.
- Applies to nonlinear models and interactions, not just linear structural-equation coefficient products.
- Provides a principled sensitivity analysis that quantifies how much hidden confounding would overturn the result.
- Connects experimental treatment effects to substantive theories of why a treatment works, advancing mechanism testing.
- Identification rests on sequential ignorability, whose mediator-outcome part is untestable and often implausible.
- Randomizing treatment does not randomize the mediator, so even experiments do not by themselves identify the ACME.
- The single-mediator framework is strained when several intermediate variables operate jointly or sequentially.
- Estimates can be sensitive to model specification for the mediator and outcome and to treatment-mediator interaction assumptions.
Common pitfalls
This section is available to Pro members. Upgrade to Pro
Applications
This section is available to Pro members. Upgrade to Pro
Frequently asked
How does this differ from the classic Baron-Kenny mediation approach?
The Baron-Kenny product-of-coefficients method estimates an 'indirect effect' from regression coefficients but never defines that quantity causally and implicitly assumes linearity and no mediator-outcome confounding. The potential-outcomes approach of Imai and colleagues defines the average causal mediation effect explicitly in terms of counterfactuals, states the sequential-ignorability assumption needed to identify it, works for nonlinear models and treatment-mediator interactions, and supplies a sensitivity analysis. The two coincide only under restrictive linear, no-interaction, no-confounding conditions. See the linked mediation-analysis entry for the classical treatment.
If my treatment is randomized, can I just trust the mediation estimate?
No. Randomizing the treatment guarantees the first part of sequential ignorability — that treatment is unconfounded — but it does not randomize the mediator. The mediator is a post-treatment variable that subjects' own characteristics influence, so the mediator and outcome can still share unobserved confounders. Identifying the ACME requires the additional, untestable assumption that the mediator is unconfounded with the outcome given treatment and covariates. This is why even experimentalists must report a sensitivity analysis or use designs that manipulate the mediator. See the linked causal-mediation entry for the general framework.
What does the sensitivity analysis actually tell me?
Because the no-mediator-confounding assumption cannot be verified from data, the sensitivity analysis asks how strong an unobserved confounder between the mediator and outcome would have to be to drive the estimated mediation effect to zero. It parameterizes this as the correlation between the error terms of the mediator and outcome models and reports the value at which the conclusion flips. If only an implausibly large confounder could overturn the result, the mediation finding is robust; if a small one suffices, the finding is fragile. It converts an untestable assumption into a transparent, quantitative robustness check.
Sources
- 1.Imai, K., Keele, L., & Tingley, D. (2010). A General Approach to Causal Mediation Analysis. Psychological Methods, 15(4), 309–334.
- 2.Imai, K., Keele, L., Tingley, D., & Yamamoto, T. (2011). Unpacking the Black Box of Causality: Learning about Causal Mechanisms from Experimental and Observational Studies. American Political Science Review, 105(4), 765–789.
You have read it. What now?
Cite this page
ScholarGate. (2026, June 22). Causal Mediation Analysis in Politics. ScholarGate. https://scholargate.app/political-science/causal-mediation-analysis-politics